Dynamic Profile-Selection for zSlices Based Type-2 ... - IEEE Xplore

WCCI 2012 IEEE World Congress on Computational Intelligence June, 10-15, 2012 - Brisbane, Australia

FUZZ IEEE

Dynamic Profile-Selection for zSlices Based Type-2 Fuzzy Agents Controlling Multi-User Ambient Intelligent Environments Aysenur Bilgin, James Dooley, Luke Whittington, Hani Hagras, Martin Henson School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK.

Christian Wagner Horizon Digital Economy Institute, University of Nottingham, Nottingham, UK.

Abstract— Ambient Intelligence (AmI) is a vision that refers to an information technology paradigm where a physical environment is ‘aware’ of its human occupants’ presence/context and is sensitive, adaptive and responsive to their needs. Physical environments that are augmented with AmI are called Ambient Intelligent Environments (AIEs) which are deemed to be intelligent because the system should be able to recognise human occupants, reason with context and program itself to meet the occupants’ needs by learning from their behaviour [1]. However, there is a need also to deal with real-world scenarios which involve multiple users occupying a given AIE. In order to handle multi-user AIEs and control them, there is a need to have agents that are able to learn the user(s) behaviours and handle the intra and inter-user uncertainties as people have different preferences and profiles which continuously change. In this paper, we present a zSlices based type-2 fuzzy agent which employs zSlices general type-2 fuzzy systems to learn the user(s) preferences and profiles and handle the encountered intra and inter-user uncertainties. The agent will behave according to a learned user profile that is unique to an individual user or a group of users and so the profile-selection problem manifests when the set of users in an AIE changes (i.e. when people enter/ leave an AIE). The proposed agent employs a novel strategy that we call Dynamic ProfileSelection that uses a cloud-based profile repository in order to support the agent activity in multiple AIEs. To demonstrate the proposed approach, we have conducted real-world experiments on two distinct AIEs which are the intelligent apartment (iSpace) and the intelligent Classroom (iClassroom) located at the University of Essex. Keywords: type-2 fuzzy logic systems, zSlices general type-2 fuzzy systems, ambient intelligent environments, ambient intelligence.

I.

INTRODUCTION

As ubiquitous computing research progresses beyond the deployment and interconnection of devices with rudimentary software, we move towards the long envisioned goal of a physical environment that exhibits a quality of Ambient Intelligent Environments (AIEs) [1]. AIEs are characterised by their ubiquity, transparency and intelligence [1], [2], [3]. AIEs are ubiquitous because the user is surrounded by a multitude of inter-connected embedded systems, transparent because the computing equipment appears invisible to the user as it is

U.S. Government work not protected by U.S. copyright

Areej Malibari, Abdullah AlGhamdi, Mohammed J. Alhaddad, Daniyal Alghazzawi King Abdulaziz University, Jeddah, Saudi Arabia.

seamlessly integrated into the background. In addition, an AIE is intelligent because the system is able to recognise human occupants, reason with context and program itself to meet the user(s) needs by learning from their behavior [1]. The complexity associated with the number, variety and purposes of computer based artefacts leads to the need for intelligence that blends seamlessly into the infrastructure of AIEs and automatically learns to carry out everyday tasks based upon the users’ habitual behaviour(s) [4]. AIEs are rich in information, some of which is sourced in real-time from a plethora of sensing devices that sample the world around them. It is with this information that intelligent agents are able to carry out decision-making and be pro-active when taking actions that manifest back in the physical world through the control of actuators. The agents operating in AIEs should be able to act in a non-intrusive mode and should be able to learn the user(s) behavior while handling the encountered uncertainties and they should be able to represent the learnt behaviour in a transparent, easy to understand manner. In our previous research, we have presented various embedded intelligent agents to learn the individual user behaviour in a single environment where we have developed type-1 [4] and type-2 fuzzy logic agents [5] which are able to learn the user behaviour and operate over long-time durations and handle the short and long term uncertainties. However, there is a need also to deal with real-world scenarios which involve multiple users occupying AIEs. In order to handle multi-user AIEs and control them, there is a need to have agents that are able to learn the user(s) behaviours and handle the intra and inter-user uncertainties as people have different preferences and profiles which continuously change. In our previous work, we have presented zSlices general type-2 Fuzzy Logic Systems (FLSs) [6] and we have shown in [7], [8] how zSlices based type-2 FLSs could be used to model agreement and hence handle inter and intra-user uncertainties in AIEs. As part of our EU funded ATRACO, we have been investigating how to enable type-2 fuzzy-logic based agents that can cope with changing populations in multi-user AIEs. This is a non-trivial challenge, as the user-preferences for a group reflect the social constructs that exist within that

particular set of users (i.e. the preferences forr a group of four people may change if one member of the groupp were to leave). Adding further complexity to the multi-useer challenge is the multi-space challenge in which the same set of users in a different AIE will have different preferences. H Hence, this paper presents our approach to solving the multi-useer problem that is consistent across multiple AIEs. To validate tthe approach, we have deployed an intelligent agent in two “livving-labs” on the University of Essex campus which are the inteelligent apartment (iSpace) and the intelligent classroom (iClassrooom). Section II begins with the introduction of the “profileselection” problem that motivates this paper. Section III then describes our zSlices based general type-2 embbedded agent that we call a “Fuzzy Task Agent” (FTA) and its ooperation; and we present also a novel approach which we call Dynamic Profile Selection that uses a cloud-based profile repository in order to support the agent activity in multiple AIEs. Seection IV presents the trigger conditions for dynamic profile seleection. Section V introduces the application of the proposed apprroach. Section VI shows the experiments and results while the conclusions and future work are presented in Section VII. II.

THE PROFILE-SELECTION PRO OBLEM

The work to date in fuzzy-logic based intellligent agents has investigated agents that are able to perfoorm environment adaptation in response to observations throuugh sensor inputs [4], [5], [9]. The learnt user(s) behaviour moddel is based on a profile that is generated by learning the prefeerences of a user. However, as we approach real-world deplooyment of agent technologies in multi-user environments, we m must consider the user(s) requirements that govern their behaviouur. Our recent social evaluation in this area has placed realworld users into environments where our agennts are active. As expected, the desired behaviour of an AIE E is sensitive to context (such as time, location, activity, moood, etc.), but we have also identified that part of this context is iinfluenced by the social relationships that exist among occupantts. That is, as the human population of an AIE changes, there iss also a change in the expected behaviour. Our current approach, as reported in this ppaper, is to assign a unique profile to every observed combinatioon of users in an AIE. It is then up to the agent to select the ccorrect profile to match the context. For example, in an AIE where two users (user-A and user-B respectively) have been obbserved, there are possible three profiles {A, B, AB}. The prroblem that now exists is how the agent will select the correcct profile as and when the users of the AIE arrive and leave. Adding further complexity, context also ccontains location; so the same users in a different AIE will have different desired behaviour. The profile selection problem thherefore needs to account not only for the combination of users, but also the location (AIE) in which the agent is operatingg. More formally, given a set of profiles P, how shall an agent seelect the profile p (where p ∈ P) given the input vector 〈 L, U 〉 where L is the location and U is the set of users that curreently occupy that location.

III.

THE ZSLICES BASED FUZZY TASK AGENT

Amongst the fundamental ch haracteristics of intelligent embedded agents is the ability (aalso the motivation [2]) to ubiquitously and transparently bring b computing into the physical world. To achieve this, ag gents have the capacity for reasoning, planning and learning [2 2] through the effective use of physical components such as sen nsors and actuators [3]. The necessity of such devices stems fro om the fact that they enable the agent to sense the state of the en nvironment as well as to act upon it [4]. Herein, we use sensors for f the agent to perceive the environment, capture the state and d then through the use of actuators (controllers) to take acctions, which perform the learned preferences of either the individual user or a group of users encapsulated in the correspond ding profiles. The first of two sub-sections heerein presents fundamentals regarding the zSlices based generall type-2 FLSs employed in the FTA and its structure. This is fo ollowed by the second subsection that presents our “Dynamic Profile Fuzzy Task Agent” (DPFTA) that is a unique and noveel kind of FTA that exhibits dynamic profile selection. A. The Architecture of an FTA e the general concept As its name suggests, our FTA employs of a fuzzy logic system. In particullar, we use a zSlices based General Type-2 Fuzzy Logic ap pproach [6] that employs membership functions in order to model m inter-device and intradevice uncertainty within a single hu uman-centered model. A zSlice is formed by slicing a general type-2 fuzzy set in the third dimension (z) at level z . The result of this slicing action is an interval set in the third dimension d with height z . In other words, a zSlice is equivaalent to an interval type-2 fuzzy set with the exception that itss membership grade , in the third dimension is not fixed to 1; instead is equal to 1. Thus, the zSlice can be written as follows where 0 [6]: / ,

(1)

where at each value (as shown n in Figure 1a), z-Slicing creates an interval set with heigh ht and domain which ranges from to as shown in Figu ure 1b, 1 , and I is the number of zSlices (excluding ) and / .

(a) (b) Figure 1. (a) Front view of a general typee-2 set (b) Third dimension at x’ of a zSlices-based generaal type-2 fuzzy set

Additionally [6]: 0/ ,

,

(2)

where is considered as a special case with zz=0. Particularly, we will consider 1 as will not ccontribute to the crisp output of the zSlices based type-2 fuzzy logic system and it can be omitted with no effects [6]. A general type-2 fuzzy set could be seenn equivalent to the collection of an infinite number of zSlices: ∞

(3)

ype-2 fuzzy sets based on we have created seven interval ty uncertainty extracted from the amalgamated information collected by various light sensors within the AIE. By the use of agreement modelling [8], we have created c an adaptive humancentred model of ambient light leevel which allows for the continuous integration of short, med dium and long term changes in the device and environment charaacteristics. We will illustrate the use of fuzzzy sets with an example in which sensor values are used to construct a human-centred model of ambient light level within an AIE. The horizontal axis of each individual fuzzy set in Fig gure 3 shows the raw lightsensor values from 800 to 1000 for the t label ‘Very High’.

In a discrete universe of discourse Equaation (3) can be rewritten as follows: (4) Throughout the paper, we will be referrinng to the discrete version in Equation (4) where the summatioon sign does not denote arithmetic addition but it denotes the unnion set theoretic operation. We have employed the max operaation to represent at of the the union; hence the membership function zSlice based general type-2 fuzzy set can be expressed as follows: / where 0

,

0,1

(a)

(b)

(5) (c)

.

The operations on zSlices based general tyype-2 fuzzy sets, namely the intersection and union operatioons implemented through the meet and join operations have beenn described in [6] and will not be the focus of this paper. Also, m more information about the details on how the uncertainty moddels are extracted from data and how the fuzzy sets are created uusing the concept of Multi-Level Agreement (MLA) could be fouund in [7][8]. The establishment of zSlices based Ambiient Light Model employs the sliding window approach shown iin Figure 2. Over the period of interest that consists of seven cconsecutive days,

(d)

(f) (e) Figure 3. Ambient Light Level Mo odel ‘Very High’ for iSpace (a)Day 1 (b)Day 2 (c)Day 4 (d))Day 5 (e)Day 6 (f)Day 7

Figure 2. Sliding window based continuous adaptation

The zSlices that are depicted in different colours across the third dimension of the model help reveal the concept of continuous adaptation by showing a potential change in the shape of the actual fuzzy set between 28/03/2011 and 03/04/2011.

clearly be noted that the learning and adaptation component acts upon the profile either to add or to modify the preferences of the individual user or the group of users. Consequently, this real-time modification drives the FTA in accordance with the “user is king” axiom [4] that has guided our research.

Figure 4 shows the internal structure of an FTA and its main functional components. Simply, the FTA processes inputs from the user (interaction through a graphical user interface) and light-sensors, and output is performed by controlling the intensity (in the range 0…100) of dimmable ceiling-lights.

Formally, the procedure of rule addition in the learning and adaptation component involves the fundamentals of fuzzy rule extraction [5]. It employs a one-pass technique that describes the relationship between , … , and ,…, to create a fuzzy rule in the following form:

For clarity, we will divide the overall architecture into two main parts according to the type of input: user input and sensory input as stated above. The user input is communicated to the FTA through an interactive environment control GUI. This expresses the users desired output – for example, if the user wants the lights on, they will touch a graphical “switch” that communicates this desire to the FTA. The consequence of this interaction is either a rule addition or a rule modification within the FTA (in addition, of course, to the manifestation of the desire through actuators). In other words, the user interaction causes a change in the profile of the user or the group of users within the AIE. All of the updates performed in the user profiles are online and real-time indicating how adaptive and responsive the agent is.

IF

is

THEN

… and is

is

… and

(6)

is

There are fuzzy sets each of which has zSlices defined for every single input where 1, … , fuzzy sets 1, … , and 1, … . Similarly, there are defined for every single each of which has zSlices output where 1, … , 1, … , and 1, … . It is important to note that the number of zSlices is the same for all the input sets, which are created in a novel approach as mentioned earlier, and output sets, which are created according to the expert knowledge. Herein, we will briefly show the steps of our approach where we have applied modifications to address the requirements of a zSlices based general type-2 fuzzy logic system as follows: Step 1) For a fixed input-output pair ; where is the desired is the snapshot of the environment and environment state chosen by the user, we compute the upper and lower membership values and for each zSlice where 1, … , for each 1, … , . We fuzzy set 1, … , and for each input find 1, … such that 1, …

, for all where

is the z-weighted center of gravity of the

membership of

at

calculated by using the following:

∑

∑ ∑

Figure 4. The FTA Architecture

Both user and sensory inputs initially pass through the fuzzification process where the fuzzy membership functions are represented by zSlices based general type-2 fuzzy sets. Together with a snapshot of sensory data that models the AIE state, user input goes through a series of actions that allows the learning and adaptation component of the FTA to reconcile the rule base (human interpretably-called profile). Particularly, the user input occurring in a specific environment state is checked against the rule base. If the profile includes this current preference, adaptation takes place (rule modification) whereas learning occurs otherwise (rule creation / addition). It should

(7)

∑

(8)

/ and 1 . The antecedents generate the where IF part of the rule as in (6) after Step 1 is completed for each input . The linguistic labels selected to form the first part of the extracted rule can be represented as follows: IF

is

… and

(9)

is

Step 2) The calculations (7) and (8) are applied on the with the corresponding fuzzy set . Hence, output (7) and (8) can be rewritten as (10) and (11) successively: , for all

1, …

(10)

∑

∑ ∑

∑

(11)

For simplicity, we will show only one output for the consequents of the generated rule which has the following final form: IF

is

… and

is

THEN

is

(12)

Sensory inputs allow the FTA to perceive changes in the AIE environment. After going under the similar fuzzification procedure as the user input, the sensory input follows the natural series of actions in a zSlices based type-2 fuzzy logic system. The fuzzy inference enables the usage of the most upto-date rule base and hence, the FTA transparently performs the learned or adapted preferences of the user on the environment. The rest of the architecture including type-reduction and defuzzification processes are naturally inherited from the structure of a zSlices based general type-2 fuzzy logic system as described in [6]. In the end of these calculations, the crisp output reflecting the preference of the user or the group of users is sent to the actuators within the AIE, hence facilitating the control of the environment on behalf of the user in an intelligent and comprehensible way. B. Dynamic Profile Fuzzy Task Agent (DPFTA) The rule base in a fuzzy logic system makes those systems uniquely human-interpretable together with the notion of linguistic variables. In our approach, the rule base forms a “profile” of either an individual user or a group of users in which corresponding preferences are kept in the form of IfThen rules. The FTA makes use of the profile to set the environment state according to learned user desires. The term “dynamic” highlights the fact that the agent can change the profile that is actively used during runtime. An FTA that has this ability to swap out the current profile for a replacement profile distinguishes it from other forms of FTA and is identified as a “Dynamic Profile Fuzzy Task Agent” (DPFTA). The combination of these properties allows a DPFTA to replace the profile for a user (or group of users) with the profile of another user (or group of users), thus changing the particular behaviour of the FTA. This in turn changes the way in which the FTA can manipulate the physical space resulting in different environmental adaptation behaviour as perceived by the user(s). Furthermore this can all be achieved at runtime. Upon the recognition of some trigger condition that is necessary to prompt a change in the active profile used by the agent, the DPFTA will execute the “changeProfile” algorithm as outlined by the pseudo-code in Figure 5. IV.

TRIGGER CONDITION AND PROFILE-SELECTION

One of the main features that an AIE is intended to exhibit is automated adaptation in response to contextual change. This requires agents in general and the FTA in particular to be context aware through the use of sensors. It is an extremely important user requirement that an AIE operates in a robust and real-time fashion. To achieve this, agents must continuously

sample the world and react to the changes that they observe. Some of these changes are subtle and proportional to the observed change of context, while other observations are more identifiable as discrete events that result in major behavioural changes. A change in the human population of an AIE is the trigger condition that prompts a DPFTA to change its active profile by invoking the “changeProfile” algorithm as defined in Figure 5. Two recognizable events are used to detect this change: 1.

The arrival of a new user,

2.

The departure of a current user.

The most significant part of the “changeProfile” operation is the “selectNewProfile” algorithm that requires access to two parameters: 1.

A set of accumulated user profiles that have been previously created and from which to perform selection,

2.

The current context of the AIE in which the DPFTA is operating.

Procedure changeProfile(Context context) Begin // given the new context, select a new profile nextProfile = selectNewProfile(context, userProfiles); // if a suitable profile was not found, then // create a new one ! if(nextProfile == null) nextProfile = new Profile(); // halt the current profile operation activeProfile.stop(); // set the new profile activeProfile = nextProfile; // start the new profile operation activeProfile.start(); End Figure 5. Pseudo-code for the changeProfile algorithm

When a profile change is triggered, the DPFTA extracts the relevant information that it needs from the current context (i.e. the identities of the AIE and its human occupants) and uses that information to select a profile (p) from the set of existing profiles (P) that are available (where p ∈ P). Should no suitable match be found, then a new default profile is created and added to P. In addition to the procedure of profile-creation, we introduce the profile-selection policy, which will facilitate the differentiation between various spaces having particular context. Practically, the policy reflects the social interaction model used by users in the associated AIE and governs the profile-selection algorithm that each DPFTA uses. Taking the context into consideration, we identify a “mode” of operation, which drives the activity of the agent in managing the environment and allows for a user-led approach. The modes that we define for the specific application will be detailed in the following section. Herein, it is important to note

that the modes are not limited and can be maanifold depending on the desired social interaction. V.

OACH APPLICATION OF THE APPRO

To demonstrate the application of our D DPFTA approach across multiple AIEs; user-profiles are stored in a cloud-based repository and facilitate the operation of lightinng control agents that are deployed in two multi-user AIEs at tthe University of Essex (the iSpace and iClassroom). Figurre 6 shows the employed architecture.

When posting to the server, the t user’s entire profile is stored in the database, even if emp pty, serving as a record of users logged into the space. Howeever, if the web service is polled for a user that does not exist in the database, rather than create an empty profile, it will reeturn HTTP code 204 (the server fulfilled the request but no co ontent was returned). The database structure takes a user-centric approach as shown in TABLE I, and indexes each profile by using a combination of user-ID and space-ID D, thus profiles are specific to a user set and location. The user-IID can refer to a single user (as is the case for users A and B in n TABLE I) or a set of users whose IDs are concatenated (as is th he case for the profile AB in TABLE I). TABLE I. USER REPOSITORY TABLES THAT T INDEX PROFILES USING USER-ID AND SPACE-ID AS S KEYS

User-ID

Space-ID

Profile

A

iSpace

A_iSpace.ser

iSpace

B_iSpace.ser

iClassroom

B_iClassroom.ser

iSpace

AB_iSpace.ser

B AB

Figure 6. The employed architecture

Each AIE contains a set of network-conneccted artefacts that are wrapped using UPnP to hide the heterogenneous technology / communications that underpin the constructtion of the space. This approach has been widely used in our preevious works [10] and provides a common homogeneous interfaace for a DPFTA to manipulate geographically distributed inpuuts / outputs that manifest in the real world. Of those artefactss deployed in our AIEs, the following are used in this work: 1.

Dimmable lights that are embedded inn the ceiling and used as outputs to be controlled by the DPFTA.

2.

Wall / Ceiling mounted light-level seensors that act as inputs to the DPFTA.

3.

A “FollowMe” GUI [11] that allowss human users to provide input to the DPFTA (i.e. ask fo for the light levels to be modified, and set the profile-seleection policy).

4.

RFID readers that allow users to “touuch-in / out” of a space, thus allowing the DPFTA to m monitor occupant population.

It is possible for certain users to have more spaces than others, likewise, it is possible for spaces to have more users than others. The profiles can be acceessed either via the browser (using a parameterized URL) or through t client code (which creates the appropriate HTTP GET G and POST requests necessary). VI.

EXPERIMENTATIION & RESULTS

Several experiments have been n conducted with up to 8 participants, each having a differentt RFID tag, in two different AIEs (iSpace and iClassroom). Fig gure 7 shows photos of the users conducting experiments in the iClassroom and the iSpace. Classroom is different from Practically, the context of the iC the context of the iSpace and hen nce, we define two profileselection policy modes: “master mode” m and “peer mode”. In order to maintain a reasonable lev vel of complexity, we have opted to use the peer mode as the deefault in both spaces.

Both AIEs are connected to the Univversity of Essex research network, through which they can both access the profile repository that is deployed elsewhere inn the University.

When in the peer mode, a DPFTA D creates profiles for individual users whenever they are present in the space alone. But when multiple users are presentt, profiles are created for the group. However, when the DPFTA A is set to operate in the master mode, users have an associaated priority and the profile that belongs to the single highest priority p user in the space at that time will be selected. This iss useful, for example, in a classroom scenario where the teacher holds a higher priority over the students and so is enforced d. When the teacher leaves, the next highest priority user is seleected. If one does not exist, then the system behaves as in the peeer mode.

The “User Profile Repository” (UPR) is a Tomcat based web-server equipped with JAX-RS that is tiied to a MySQL database, allowing deployment of a RESTful web service. As each user interacts with the AIE, a profile is generated by the DPFTA and uploaded to the UPR for storage.

The results of our experiments can be visualized from the profiles generated by the DPFTA an nd stored in the database on the server. Herein, we will illustratte a small portion of one of the several scenarios, and show the actual rule bases as well as the profiles stored in the databasse whenever a participant

arrives and leaves an AIE. It is important to note that the timeline (history) is essential in the experriments, in other words, as adaptation takes place over time, thhe events that are happening for the first time should be distingguished from the events that have already happened before. Foor simplicity, we will start detailing the results from the beginninng in a sequential manner.

IF time of day is AFTERNO OON and Ambient Light Level is LOW THEN Ceiling Lights is HIGH The output of the DPFTA befo ore the arrival of Peer_2 in iClassroom is illustrated in Figure 8. DPFTA EXPERIMENT @ iClassroom WELCOME peer_1 ---RULEBASE CREATION--- for peer_1 Rulebase for peer_1 BEFORE ADAPTATION is EMPTY! ---END OF RULEBASE CREATION--- for pe eer_1 ---ADAPTED (AFTER ADAPTATION) RULE EBASE--- for peer_1 If Time of Day is AFTERNOON and Ambient Light Level is LOW then Ceiling Lights is HIGH ---END OF ADAPTED RULEBASE--- for pe eer_1

Figure 8. The system output of the DPFTA A for peer_1 before the arrival of peer_2 in the iClaassroom

The cloud-based repository is updated u as shown in TABLE III. The new profile is created forr the group of Peer_1 and Peer_2 together. The adaptation takees place in the group profile without affecting the individual profile of Peer_1, which is copied into the group profile Peeer_1Peer_2 and used as a starting point that the users can mod dify as they see fit. (a)

TABLE III. USER-ID AND SPACE-ID COL LUMNS OF THE UPR DATABASE AFTER ARRIVAL OF PEER_2 TO O THE ICLASSROOM

(b) Figure 7. Three of the users carrying out adaptatioon experiments in the (a) iClassroom (b) iSpace

When the participant Peer_1 arrives in thee iClassroom and touches in with his / her RFID tag, the DP PFTA creates an empty rule base (i.e. an empty individual proofile). Unless any other peer or the master user arrives at the sppace, the DPFTA adapts to the preferences of Peer_1. TABLE III shows the UPR User-ID and Space-ID columns of the databasee table. TABLE II. USER-ID AND SPACE-ID COLUMNS OF THE E UPR DATABASE AFTER ARRIVAL OF PEER_1 TO THE ICLAS SSROOM User-ID

Space-ID

Peer_1

iClassroom

Peer_1 chooses to use the GUI to set the llights to bright in the iClassroom. This is communicated to the DPFTA and the profile is augmented with the following IF-TH HEN rule:

User-ID

Space-ID

Peer_1

iClassroom

Peer_1Peer_2

iClassroom

The peer mode is the default mode m of operation for the DPFTA and is valid unless the masster user (if defined) arrives at the space. Upon arrival of the teacher in the iClassroom, who is introduced as the master user in this case, the DPFTA creates a new profile only for Teeacher. The control of the environment is handled by Teacher and any interaction through the GUI allows for the adaaptation of this profile only. At this point, it is important to notee that there are already two peers inside the iClassroom, Peer_1 1 and Peer_2, whose group profile Peer_1Peer_2 is saved and sent to the UPR. The scenario continues with the following steps that also include another space (the iSpace – a fully furnished apartment AIE). TABLE IV shows the UPR contents c after the following steps are complete. • • • • •

Teacher leaves the iClassroo om, Peer_4 enters the iSpace, m and enters the iSpace, Peer_3 leaves the iClassroom Peer_5 briefly enters the iSpace and then leaves, Peer_6 arrives at the iSpace.

PR that Peer_3 has changed It can be observed from the UP his / her location and the default mode of operation of the DPFTA allows for a new profile creeation taking the contextual information into account. Also, th he first part of the system output of the DPFTA before the t arrival of Peer_5 is demonstrated in Figure 9.

TABLE IV. USER-ID AND SPACE-ID COLUMNS OF THE UPR DATABASE AFTER EXPERIMENT CONCLUDES

User-ID

Space-ID

Peer 1

iClassroom

Peer_1Peer_2

iClassroom

Teacher

iClassroom

Peer_1Peer_2Peer_3

iClassroom

Peer_4

iSpace

Peer_3Peer_4

iSpace

Peer_3Peer_4Peer_5

iSpace

Peer_3Peer_4Peer_6

iSpace

It is important to realize that the user preferences and the state of the AIE differ with regards to context.

out across two geographically separated AIE spaces on the University of Essex Colchester Campus. We have also presented the system output of the DPFTA from our real-world experiments. We have also pointed out that the preferences and the adaptation of user behaviour models depend on the context in a multi-space scenario. As part of future work, we aim to optimize the response time of the server and the size of the rule base files, which currently operate in a sub-optimal fashion. These are however implementation tweaks; of a greater importance and interest are the possibilities to define relationships between the profiles in order to assist the adaptation of the FTA. Furthermore, we seek to investigate the application of real-time profile merging that is socially acceptable are among our future investigation, thus the change in user behaviour whilst in one space may influence the behaviour of another space without re-learning. ACKNOWLEDGMENT

DPFTA EXPERIMENT @ iSpace WELCOME peer_4 ---RULEBASE CREATION--- for peer_4 Rulebase for peer_4 BEFORE ADAPTATION is EMPTY! ---END OF RULEBASE CREATION--- for peer_4 ---ADAPTED (AFTER ADAPTATION) RULEBASE--- for peer_4 If Time of Day is AFTERNOON and Ambient Light Level is LOW then Ceiling Lights is MEDIUM If Time of Day is AFTERNOON and Ambient Light Level is VERY LOW then Ceiling Lights is HIGH ---END OF ADAPTED RULEBASE--- for peer_4 WELCOME peer_3 ---RULEBASE CREATION--- for peer_3peer_4 Rulebase for peer_3peer_4 has been copied from peer_4 Rulebase for peer_3peer_4 BEFORE ADAPTATION If Time of Day is AFTERNOON and Ambient Light Level is LOW then Ceiling Lights is MEDIUM If Time of Day is AFTERNOON and Ambient Light Level is VERY LOW then Ceiling Lights is HIGH End of TO-BE-ADAPTED Rulebase for peer_3peer_4 ---END OF RULEBASE CREATION--- for peer_3peer_4

Figure 9. The system output of the DPFTA before the arrival of peer_5 in the iSpace

VII. CONCLUSIONS & FUTURE WORK In this paper, we present a zSlices based type-2 fuzzy agent which employs zSlices general type-2 fuzzy systems to learn the user(s) preferences and profiles and handle the encountered intra and inter-user uncertainties. The agent will behave according to a learned user profile that is unique to an individual user or a group of users and so the profile-selection problem manifests when the set of users in an AIE changes (i.e. when people enter/ leave an AIE). The proposed agent employs a novel strategy that we call Dynamic Profile-Selection that uses a cloud-based profile repository in order to support the agent activity in multiple AIEs. This approach is motivated by the need to change the profile that a fuzzy agent uses at runtime in response to a change in the multi-user population of an AIE that the agent caters to. We have also described an architecture that enables this approach to use a common profile repository; the approach therefore supports the operation of multiple AIEs and the set of users that roam between them. We have presented the results of experimentation that have been carried

This work has been carried out as part of the ScaleUp project with funding from King Abdulaziz University, and the ATRACO project with funding from the EU FP7. REFERENCES [1]

W. Weber, J. Rabaey, and E. Aarts, Ambient Intelligence. The Netherlands: Springer-Verlag Berlin Heidelberg, 2005. [2] M. Coen, “Design Principles for Intelligent Environments,” in Proceedings of the 15th National Conference on Artificial Intelligence, AAAI Press, Menlo Park, California, pp. 547-554, 1998. [3] D. Cook, and S. Das, “How smart are our environments? An updated look at the state of the art,” Pervasive and Mobile Computing, vol. 3, no. 2, pp. 53-73, 2007. [4] H. Hagras, V. Callaghan, M. Colley, G. Clarke, A. Pounds-Cornish, and H. Duman, “Creating an ambient-intelligence environment using embedded agents,” IEEE Intelligent Systems, vol. 19, no. 6, pp. 12– 20, 2004. [5] H. Hagras, F. Doctor, V. Callaghan, and A. Lopez, “An incremental adaptive life long learning approach for type-2 fuzzy embedded agents in ambient intelligent environments,” IEEE Transactions on Fuzzy Systems, vol. 15, no. 1, pp. 41-55, 2007. [6] C. Wagner, and H. Hagras, “Toward general type-2 fuzzy logic systems based on zslices,” IEEE Transactions on Fuzzy Systems, vol. 18, no. 4, pp. 637-660, 2010. [7] C. Wagner, and H. Hagras, “Employing zSlices based general type-2 fuzzy sets to model multi level agreement,” in Proceedings of the 2011 IEEE Symposium on Advances in Type-2 Fuzzy Logic Systems (T2FUZZ), pp. 50-57, April 2011. [8] C. Wagner, and H. Hagras, “An approach for the generation and adaptation of zSlices based general type-2 fuzzy sets from interval type2 fuzzy sets to model agreement with application to Intelligent Environments,” in Proceedings of the IEEE International Conference of Fuzzy Systems, Barcelona, Spain, pp. 1-8, July 2010. [9] U. Rutishauser, J. Joller, and R. Douglas, “Control and Learning of Ambience by an Intelligent Building,” IEEE Transactions on Systems, Man & Cybernetics, vol. 35, no. 1, pp. 121–132, January 2005. [10] A. Holmes, H. Duman, and A. Pounds-Cornish, “The iDorm: Gateway to Heterogeneous Networking Environments,” in Proceedings of the International ITEA Workshop on Virtual Home Environments, Paderborn, Germany, February 2002. [11] J. Dooley, C. Wagner, H. Hagras, and G. Pruvost, “FollowMe: The Persistent GUI,” in Proccedings of the 1st International Workshop on Situated Computing for Pervasive Environments (SCOPE’11) at the 6th International Symposium on Parallel Computing in Electrical Engineering (PARELEC’11), Bedfordshire, UK, 2011.